Transfer Learning Based Retinal Layer and Fluid Segmentation using Fully Convolutional Networks

Automatic segmentation of retinal structures in optical coherence tomography (OCT) b-scan images is still an open problem, especially when there is a variation in the normal structure of the retina due to the presence of diseases. This line of research has shown enormous achievements with the introduction of deep learning algorithms. This paper presents a comparative study of pre-trained fully convolutional networks for end-to-end segmentation of retinal boundaries and fluid pockets. The models are trained using a novel publicly available dataset of 1136 annotated OCT b-scans with segmentation masks for eight retinal regions. The dataset was assembled using eye images from patients with neovascular age-related macular degeneration. Empirical results show an improvement in the state-of-the-art segmentation, and near ophthalmologist level performance, with a mean dice score of 84.6% for the best model.